[ntp:questions] NTP 4.2.2p4 loopstats

David L. Mills mills at udel.edu
Tue Nov 14 04:51:47 UTC 2006


TomD,

My apologies; I did change the jitter and noise statistics quite a while 
ago, but the interval between numbered versions overlapped the change.

If you were making serious assumptions about sys_jitter and 
clock_jitter, be advised sys_jitter is not an effective diagnostic 
statistic. It is derived from the weighted average of the survivors 
jitter and the selection jitter (ntpq rv jitter). But, this is not 
indicative of the actual jitter seen by the discipline loop (ntpq rv 
noise), which is a much more sensitive indicator of loop dynamic 
performance. In addition, up until the time of the change, the jitter 
reported by the daemon and by the kernel were calculated differently. 
Now they are the same.

Dave

TomD wrote:

> Dave,
> 
> re: "the clock discipline jitter has been separated from the selection
> jitter"
> & "loopstats have not changed in many years"
> & "This is the same data reported over two decades"
> 
> No problem with separating the two jitters - but it was a surprise that
> the loopstats jitter field  switched to reporting only the new
> clock_jitter, rather than continuing to report sys_jitter starting with
> NTP 4.2.2.
> 
> clock_jitter appears to be a 3rd derivative vs. sys_jitter which
> appears to be a 2nd derivative of the clock offset.  Is that correct?
> 
> Graphing the jitter value from loopstats shows a much smaller value
> starting with 4.2.2.  It doesn't correspond clearly to real-world
> things like network load, as it did before the change. (although I
> haven't collected enough of the new 'jitter' values to be sure of this)
> 
> -tom-
> 
> p.s. This topic seems to have gotten out of hand.  The original
> question was "did jitter in loopstats change" and "what's a simple
> definition of the new value".
> 
> Well, at least we've answered jg's first question:  'Yes, but the old
> value is still available with a small code change.'
> 
> I'm a little taken aback by all the other things this topic has
> meandered into: e.g. *what values ntpq displays, *what it calls them,
> *the ntpq op command,  *the ntpq opeers value (bug 666), *ntpdc,
> *ntp.isc->udel download links, *the other loopstats fields, *projects
> to design new analysis mechanisms, etc. etc. etc.  Yikes!
> 




More information about the questions mailing list